The Reclassification: How AI Became Infrastructure at JPMorgan

The most important AI story of the week is not a model release. It is an accounting decision.

This year JPMorgan Chase set its technology budget at roughly $19.8 billion, up about $2 billion from the prior year. The headline number got the coverage it always does — the biggest bank in America spends more on technology than most countries spend on anything. But the number is not the story. The story is where the money moved from and to. The bank took its roughly $2 billion in annual AI spending and moved it out of the experimental innovation budget — the place companies park bets they might lose — and into the same category as payment systems and data centers. The same category as the rails that move trillions of dollars a day and cannot be switched off without the institution failing.

That is the whole story. Not the dollar figure. The reclassification.

When you move a line item from "innovation" to "infrastructure," you are not making a bigger bet. You are making a different kind of statement. Innovation spending is discretionary by definition: it is the budget you cut first when the quarter turns ugly, the portfolio of options where most are expected to expire worthless. Infrastructure spending is the opposite. It is non-negotiable. It is the plumbing. You do not run a sensitivity analysis on whether to keep the payment network on. The reclassification is JPMorgan declaring, in the only language a bank truly speaks — the budget — that generative AI has crossed from the column of things it is trying into the column of things it runs on.

What makes this remarkable is that the bank made the move while openly admitting it cannot prove the return. That tension — infrastructure you can't turn off, returns you can't yet measure — is the actual subject of this piece.

What actually happened

Strip away the framing and here is the concrete sequence.

The bank did not announce a moonshot. It did the unglamorous thing: it changed which ledger the spending lives on. Generative AI's share of the bank's overall AI activity is rising, building on a base of AI use cases that roughly doubled in production through 2025. The newly earmarked $1.2 billion is pointed at three high-impact, decidedly operational areas — automating customer-service work in call centers, generating personalized insights for clients, and building tools that make the bank's own software engineers faster. An internal coding assistant has already been credited with improving engineer productivity by as much as 20 percent.

None of those line items reads like research. They read like operations. And operational spending that recurs every year, that the business depends on to function, belongs — by the plain logic of corporate accounting — in infrastructure, not in the experimental fund. The reclassification is less a strategic gamble than a bookkeeping admission that the experiment already ended.

Why "core infrastructure" is the entire point

To see why the category matters more than the dollars, you have to look at what each budget classification actually does inside a large company.

A line in the innovation budget is a hypothesis with a dollar sign attached. A line in the infrastructure budget is a dependency. The difference is not cosmetic — it changes who defends the spend in a downturn, how it is governed, and what happens to it when returns are slow to materialize. An innovation bet that underperforms gets euthanized. An infrastructure dependency that underperforms gets fixed, because turning it off is not on the table.

That is why the reclassification is a stronger signal than any spending figure. A bank could triple its AI innovation budget and still be saying "we are making a large bet we might lose." By moving the money to infrastructure, JPMorgan is saying something categorically different: we have already decided this is load-bearing, and we are now going to maintain it like the payment network rather than evaluate it like a startup. Conviction is cheap in press releases. It is expensive — and therefore credible — in the chart of accounts.

The numbers behind the move

The reclassification did not happen in a vacuum. It is the punctuation mark at the end of a multi-year ramp.

The technology line has climbed steadily for years, but the composition inside it has shifted faster than the total. AI went from a rounding error to a roughly $2 billion recurring commitment, and the most recent increase is disproportionately AI-weighted. Of the roughly $2 billion year-over-year rise in the total budget, more than half — that $1.2 billion — is explicitly tagged to AI high-impact areas.

The precise internal split is not public; the chart above is an illustrative breakdown consistent with the three priority areas the bank has named. What is not illustrative is the direction: the money is flowing toward functions that employ large numbers of people doing repeatable, language-shaped work — exactly the surface that current agentic systems are best at absorbing. Call-center operations and back-office processing are not chosen at random. They are chosen because they are where the measurable cost sits.

The deeper change is visible only if you plot the spend by which budget it lives in over time.

That migration — warm experimental dollars cooling and consolidating into one dense, permanent infrastructure line — is the same shape as the hero image above this article, and it is not a coincidence. It is the literal picture of what "core infrastructure" means: many scattered, optional bets compressed into a single load-bearing column the institution now depends on.

The ROI paradox

Here is the part that should make you sit up. JPMorgan reclassified AI as infrastructure while its own leadership was on the record saying the returns are hard to measure. Jamie Dimon has long described technology returns as elusive — not absent, but genuinely difficult to pin to a number, the way it is genuinely difficult to attribute a quarter's revenue to the existence of the electrical grid. The bank credits its internal coding assistant with up to a 20 percent engineering productivity gain, but even that figure carries the usual caveats of self-reported developer productivity, a metric notorious for resisting clean measurement.

So the bank is in a strange posture: high enough conviction to treat AI as non-discretionary plumbing, but low enough measurement confidence to admit it cannot cleanly prove the payback. Plot those two axes against each other and the reclassification lands in a quadrant that, historically, has been rare and consequential.

How does a famously disciplined institution justify funding something as non-negotiable that it cannot prove pays for itself? The honest answer is that it is not justifying it as a return on investment at all. It is justifying it as a cost of remaining a bank. This is the logic of infrastructure: you do not build a fraud-detection system because you can prove its incremental NPV to three decimal places. You build it because operating without it is not a coherent option. JPMorgan has decided that operating without pervasive AI is, on a five-year horizon, not a coherent option either — and once you believe that, the ROI question quietly changes from "does this earn its keep?" to "what does it cost us to not have it?"

That reframing is the most important thing happening in enterprise AI right now, and it is happening on a spreadsheet, not a benchmark. It is the demand-side mirror of the supply-side bubble debate we examined in the analysis of Anthropic's $965 billion S-1: the vendors are raising capital against a future of pervasive enterprise demand, and here is one of the largest enterprises on earth booking that demand as permanent before the returns are even legible.

The signal is the institution, not the spend

If a venture-backed startup reclassified AI as core infrastructure, no one would blink — startups are supposed to bet the company. The signal here is precisely that it is JPMorgan: a 165-year-old, capital-constrained, regulator-scrutinized, risk-committee-governed institution whose entire culture is built around not treating unproven things as load-bearing. When that kind of organization moves a line item into the can't-cut column, it is making a statement other large enterprises read carefully, because banks are professional skeptics about their own spending.

The herd effect here is structural, not faddish. Enterprise software buying is deeply imitative: large incumbents watch what the most respected peer in their sector does and use it as cover. "JPMorgan treats this as infrastructure" is the single most useful sentence a CIO at a smaller bank can put in a budget memo, because it converts an unprovable bet into a defensible peer benchmark. The same dynamic propagated cloud adoption, then zero-trust security, then — over the past year — the Model Context Protocol becoming the de-facto integration standard for enterprise AI. Reclassification is contagious precisely because the accounting category, unlike the technology, requires no technical conviction to copy.

This is also why the move connects to the procurement story we covered in the collapse of AI procurement into universal credits. Once AI is infrastructure, it gets bought like infrastructure — in large, committed, multi-year capacity deals rather than per-seat experiments. The accounting reclassification and the procurement reclassification are the same phenomenon viewed from two sides of the same ledger.

Where the people are

There is no honest version of this article that skips the displacement question, because the three areas JPMorgan named — call centers, client servicing, software engineering support — are not abstract. They are staffed by people. The reason call-center automation tops the priority list is the same reason it tops every enterprise AI roadmap: it is a large, measurable, language-shaped cost center, which is exactly the profile current agentic systems compress most effectively.

Notice the bottom of that chart. The functions least exposed are the ones where a credentialed or accountable human must own the result — the risk sign-off, the compliance attestation, the regulated advice. That is the same structural moat we traced across professions in the analysis of why persistent-memory agents are absorbing customer-support roles: where the law or the regulator forces a named human to be accountable, seats survive; where the work is delegated and unaccountable, the agent absorbs it. A bank is a near-perfect laboratory for this thesis, because it contains both kinds of work under one roof, and the reclassification budget is pointed squarely at the unprotected half.

The uncomfortable arithmetic is that the same $1.2 billion that buys the productivity gain is, in part, the headcount it replaces, recapitalized as software. The "20 percent more productive engineer" is a real gain and a real reduction in future hiring at the same time. Both things are true. Treating AI as infrastructure makes that trade-off permanent rather than provisional, which is exactly why the accounting move matters to workers as much as to investors.

The bear case

It would be irresponsible to present the reclassification as self-evidently correct. There is a coherent bear case, and it is worth stating plainly.

The first risk is that "infrastructure" becomes a place to hide unaccountable spending. The entire value of the innovation budget is that it forces a kill-or-scale decision; once you move spend to infrastructure, it stops getting that scrutiny. A cynic would say JPMorgan did not reclassify AI because it is proven, but because reclassifying it exempts it from having to be proven. That is a real failure mode, and the only defense against it is the disclosure of hard numbers — which the bank has not yet fully provided.

The second risk is durability of the cost curve. Infrastructure assumptions depend on stable unit economics. If inference prices or model-licensing terms move sharply — in either direction — the "permanent line item" could prove far less stable than payment-network costs, which barely move from year to year. AI is being booked with the permanence of a data center while still carrying the price volatility of a frontier technology.

The third risk is the most basic: the productivity gains may be real but insufficient to justify the absolute spend, and "we can't measure it" can mask "it doesn't pay." The reclassification does not resolve that question. It postpones it — which is fine if the conviction is right and expensive if it is not.

What to watch

The reclassification turns a fuzzy narrative into something close to falsifiable, which is the part I find genuinely useful. Here is what would confirm or undercut the thesis over the next 18 months.

• Disclosure of hard numbers. The strongest confirmation would be JPMorgan, or a peer, publishing a specific, quantified productivity or cost-savings figure attributed to AI in an official filing or earnings call — not a vibe, a number. I have written this up as a concrete, dated prediction that the largest US banks will disclose quantified AI savings by the end of 2027.
• Peer reclassification. Watch whether Bank of America, Citi, Wells Fargo, and Goldman move AI spend into core-technology or infrastructure categories in their own disclosures. The herd signature would be unmistakable.
• Headcount composition. Watch the ratio of engineers-to-output and support-staff-to-customer, not just absolute headcount. Reclassification predicts the denominator shrinking even if the top line grows.
• The first downturn test. The real proof of "infrastructure" is what survives a bad quarter. If AI spend gets cut alongside discretionary projects the next time earnings disappoint, the reclassification was rhetorical. If it is defended like the payment network, it was real.

That last signpost is the one I will be watching hardest. Anyone can call something infrastructure in a good year. The category only means what it claims to mean when the cycle turns and the institution has to choose what it cannot live without.

The takeaway

The lesson of the JPMorgan reclassification is not that AI has "won" or that the ROI is proven — it explicitly is not. The lesson is subtler and more durable: the most serious enterprises are no longer treating AI as a thing to evaluate. They are treating it as a thing to depend on, and they are doing so before the spreadsheet can fully justify it, the way every genuine infrastructure decision in history has been made on conviction ahead of proof.

When the most skeptical, most regulated, most discipline-obsessed institution in American finance moves a $2 billion line out of the column where bets go to die and into the column where the lights live, it is telling you what it believes about the next decade in the only language that costs it anything to be wrong about. The headline number is a distraction. Watch the category. The category is the conviction.

Further reading

• Anthropic's $965 billion S-1 and the bubble test — the supply-side mirror of this demand-side story.
• The collapse of AI procurement into universal credits — what happens to buying once AI becomes infrastructure.
• Why persistent-memory agents are absorbing customer-support roles — the displacement thesis behind the call-center earmark.
• Prediction: the largest US banks will disclose quantified AI savings by end of 2027 — the falsifiable follow-through.